Summary

  • Article 49 requires a high-risk AI system listed in Annex III to be registered in the EU database before it goes on the market, and Annex VIII prints the thirteen fields of that public entry. Not one of them is an incident field.
  • Article 73 obliges the provider to report a serious incident to the market surveillance authorities 'immediately' after establishing a causal link or its reasonable likelihood, with 15 days as the outer bound, two days in the gravest cases and 10 days where a person died. No provision reviewed on 18 November 2026 requires that report to be published.

Start with the form. Before a high-risk AI system is placed on the market or put into service, Article 49 requires the provider — or, where applicable, the authorised representative — to register itself and the system in the EU database established by Article 71 (Article 49). Annex III point 2 systems are carved out and registered nationally instead. Everything else walks through the same door.

Annex VIII then prints what the entry contains, and the list is worth reading in full, because it is the specification of what Europe has decided a stranger needs to know about a machine that will decide things about people. Section A carries thirteen items: the provider's identity and contact details; the identity of the person submitting on its behalf; the authorised representative; the trade name and traceability reference; the intended purpose and the supported components or functions; a basic description of the data, inputs and operating logic; the system's status; notified-body certificate details; a scanned certificate; the Member States where the system is available; the EU declaration of conformity; the electronic instructions for use, except for Annex III points 1, 6 and 7; and an optional additional-information URL (Annex VIII). Where a public authority deploys the system, Section C adds five more: the deployer's identity and contact details, the person submitting on its behalf, the URL of the provider's entry, a summary of the fundamental-rights impact assessment findings, and where applicable a summary of the data protection impact assessment.

Thirteen fields describe a capability: who built it, what it is for, what it was trained to look at, whether a notified body signed it off, which markets it entered. A stranger can audit every one of them against a document. None of them asks what the system did to anyone after it went live. There is no incident field in Section A, Section B or Section C as retrieved, and the annex requires the information to be provided and thereafter kept up to date — a maintenance duty over a description, not over a performance history.

Article 71(4) is the sentence that makes the register genuinely public, and it is unusually direct about the standard: 'With the exception of the section referred to in Article 49(4) and Article 60(4), point (c), the information contained in the EU database registered in accordance with Article 49 shall be accessible and publicly available in a user-friendly manner. The information should be easily navigable and machine-readable' (Article 71). Two doors stay shut: the secure non-public section for Annex III points 1, 6 and 7 — law enforcement, migration, asylum and border control — where access is limited to the Commission and the authorities named in Article 74(8), and the real-world testing information registered under Article 60, which is visible only to market surveillance authorities and the Commission unless the provider consents. The default is publication, and it is publication of a description.

Now turn the page. The Act's provision about a system that has hurt someone runs the other way, to a different reader, and it starts with a different verb.

Article 73(1): 'Providers of high-risk AI systems placed on the Union market shall report any serious incident to the market surveillance authorities of the Member States where that incident occurred' (Article 73). The addressee is national and territorial. Not the Commission. Not the EU AI Office. The authority of the Member State where it happened — which, under Article 74, may be a sectoral product regulator for Annex I systems, a financial supervisor for financial institutions, or the data protection supervisory authority for several Annex III categories.

What counts as a serious incident is defined, not left to taste. Article 3, point (49) gives four limbs: death of a person or serious harm to a person's health; serious and irreversible disruption of the management or operation of critical infrastructure; infringement of obligations under Union law intended to protect fundamental rights; serious harm to property or the environment (Article 3). Limb (b) is the one that later carries the shortest clock.

Article 73 is often summarised as a fifteen-day rule. Read the paragraph, and the number is an outer bound beneath a much harder verb. The report is due 'immediately after the provider has established a causal link between the AI system and the serious incident or the reasonable likelihood of such a link, and, in any event, not later than 15 days after the provider or, where applicable, the deployer, becomes aware of the serious incident.' A second subparagraph requires the period to take account of the severity of the incident.

Two days — 'immediately, and not later than two days after' awareness — applies to a widespread infringement or an Article 3(49)(b) critical-infrastructure incident. Where a person died, the report is due immediately after a causal relationship is established 'or as soon as it suspects' one, and not later than 10 days after awareness. Article 73(5) then permits an initial report that is incomplete where that is necessary to ensure timely reporting, followed by a complete one.

That permission is the most human line in the whole instrument. It says: tell us before you have finished finding out. Article 73(6) pairs it with a duty to investigate, assess risk and take corrective action without delay, and forbids altering the system before informing the competent authorities — a rule that protects the evidence at the cost of leaving a defective system in place a little longer.

Then the machinery moves without the public. Article 73(7) requires the market surveillance authority, on receiving a notification of an Article 3(49)(c) fundamental-rights incident, to inform the national public authorities or bodies referred to in Article 77(1), and sets the Commission a guidance deadline: 'That guidance shall be issued by 2 August 2025, and shall be assessed regularly.' Article 73(8) requires the authority to take appropriate measures as provided for in Article 19 of Regulation (EU) 2019/1020 within seven days of receipt, and Article 73(11) requires national competent authorities to notify the Commission immediately in accordance with Article 20 of that Regulation (Article 73, reproduced). I did not read Articles 19 and 20 of Regulation (EU) 2019/1020, so this article says nothing about what those measures are or what that information system contains.

There is a second channel in the Act, and it is a different instrument in every respect except the hazard. Article 55(1)(c) requires providers of general-purpose AI models with systemic risk to 'keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them' (Article 55; the Commission's own page for that article carried only a non-binding summary, so the operative wording here is from a third-party mirror of the published text). No interval is printed at all — 'without undue delay' is the whole clock. And Article 55(3) routes what is obtained under that article into Article 78, whose first paragraph obliges the Commission, market surveillance authorities, notified bodies and anyone else applying the Regulation to respect the confidentiality of the information and data they receive, protecting in particular intellectual property rights and confidential business information or trade secrets, including source code (Article 78; third-party reproduction). The Act supplies the confidentiality rule itself. It does not supply a publication rule to match.

A deployer who sees the harm is a relay, not a publisher. Article 26(5) requires deployers to monitor operation on the instructions for use; where use in accordance with those instructions may present a risk under Article 79(1), to inform the provider or distributor and the relevant market surveillance authority without undue delay and suspend use; and where a serious incident has been identified, to inform 'first the provider, and then the importer or distributor and the relevant market surveillance authorities.' Only if the deployer cannot reach the provider does Article 73 apply mutatis mutandis (Regulation (EU) 2024/1689, consolidated). The natural next reader of a hospital's incident report is the vendor.

Two defects in the consolidated text are worth printing as found rather than smoothed. Article 72 requires a documented post-market monitoring system that actively and systematically collects, documents and analyses performance data across the system's lifetime — the closest thing in the Act to a duty to learn what a deployed model is doing (Article 72). But paragraph 3 has moved. The 2024 Official Journal text required a Commission implementing act establishing a template for the post-market monitoring plan and the list of elements it must contain by 2 February 2026. The consolidated text dated 27 July 2026, amended by Regulation (EU) 2026/1744, instead requires the plan to form part of the Annex IV technical documentation and Commission guidance including a template by 2 September 2027. The instrument that standardises what providers collect from the field has gone from a February 2026 implementing act to a September 2027 guidance document. Both dates are printed; neither is reconciled here.

The second defect is smaller and more telling. Annex VIII Section B points 7 and 9 are marked deleted in the consolidated text, and Article 49(4) still cross-refers to Section B point 9. A cross-reference to a field that no longer exists. That is what a register looks like when it is amended by omnibus: the structure survives the surgery, the stitches show.

Here is the boundary, in one sentence and with its reviewed set named. Across Articles 3(49), 26(5), 49, 55(1)(c), 71, 72, 73, 74 and 78 and Annex VIII as retrieved on 18 November 2026, no provision read requires the publication of an individual serious-incident report or the maintenance of a public incident register. That is an argument from silence inside those provisions, not a finding that the process is secret.

Whether the information system under Article 20 of Regulation (EU) 2019/1020 exposes any incident record publicly is not determined by the text I read; the Commission's AI Office page was requested and not retrieved, so nothing here rests on it. No figure for the number of reports received by any authority or by the AI Office appears in the ten documents retained, and none is asserted.

What this article adds beyond prior coverage, measured rather than assumed: two literal searches of this archive on 18 November 2026 — 'Artificial Intelligence Act' at limit 30 and 'serious incident' at limit 20 — returned the same two published briefings, one on Article 53 and the AI Office's hold on the 'sufficiently detailed' summary, one on the 10^25 FLOP threshold and the Commission's annex power. Both read general-purpose-model instruments about classification and disclosure. Neither reads the high-risk registration duty, the Annex VIII entry, or the Article 73 report. Against my own clock-anchored work on the European Commission — DORA's four-hour chain and the 500-kilowatt counting duty — this piece adds the artefact that carries memory, and the one that does not.

So the asymmetry is not a scandal and not a leak. It is a design decision with a price, visible in two documents built by the same legislature in the same Act. Europe publishes the machine: identity, purpose, status, certificate, market. Europe receives the incident: immediately, in two days, in ten, incomplete if necessary, with the system left unaltered until the authority has been told. The reader who consults the register to judge a deployed system is consulting a document that was never built to say what happened, and the person who was harmed is not in any field of it.

A published incident register, an anonymised periodic return, or a statistical report that names categories and outcomes would change that classification in a single amendment. Until one of them exists, the public record of European AI safety ends where the machine starts working.